Data Day Seattle, From NLP to AI

From Natural Language Processing
to Artificial Intelligence
Jonathan Mugan, PhD
@jmugan
Data Day Seattle
October 20, 2017

“ It is not my aim to surprise or shock you – but the simplest way I
can summarize is to say that there are now in the world machines
that think, that learn and that create. Moreover, their ability to do
these things is going to increase rapidly until – in a visible future –
the range of problems they can handle will be coextensive with the
range to which the human mind has been applied. ”

Herbert Simon, 1957

Herbert Simon, 1957
So what is going on here?
There is no actual
understanding.
AI has gotten smarter
• especially with deep
learning, but
• computers can’t read or
converse intelligently

This is disappointing because we want to
• Interact with our world using natural language
• Current chatbots are an embarrassment
• Have computers read all those documents out there
• So they can retrieve the best ones, answer our
questions, and summarize what is new

To understand language, computers need to
understand the world
Why can you pull a wagon with a string but not push it? -- Minsky
Why is it unusual that a gymnast competed with one leg? -- Schank
Why does it only rain outside?
If a book is on a table, and you push the table, what happens?
If Bob went to the junkyard, is he at the airport?
They need to be able to answer questions like:

Grounded Understanding
• We understand language in a way that is grounded in sensation and action.
sensation representation action
• When someone says “chicken,” we map that to our experience with chickens.
• We understand each other because we have had similar experiences.
• This is the kind of understanding that computers need.
See: Benjamin Bergen, Steven Pinker,
Mark Johnson, Jerome Feldman, and
Murray Shanahan
“chicken”

Two paths to meaning:

meaningless
tokens
manual
representations
world models
merged
representations
word2vec
seq2seq
question
answering
external world
training
symbolic
path
sub-symbolic
path

Bag-of-words representation
“The aardvark ate the zoo.” = [1,0,1, ..., 0, 1]
We can do a little better and count how often the words occur.
tf: term frequency, how often does the word occur.
“The aardvark ate the aardvark by the zoo.” = [2,0,1, ..., 0, 1]
Treat words as arbitrary symbols and look at their frequencies.
“dog bit man” will be the same as “man bit dog”
Consider a vocabulary of 50,000 words where:
• “aardvark” is position 0
• “ate” is position 2
• “zoo” is position 49,999
A bag-of-words can be a vector with 50,000 dimensions.

Give rare words a boost
We can get fancier and say that rare words are more important than
common words for characterizing documents.
Multiply each entry by a measure of how common it is in the corpus.
idf: inverse document frequency
idf( term, document) = log(num. documents / num. with term)
Called a vector space model. You can throw these
vectors into any classifier, or find similar documents
based on similar vectors.
10 documents, only 1 has “aardvark” and 5 have “zoo” and 5 have “ate”
tf-idf: tf * idf
“The aardvark ate the aardvark by the zoo.” =[4.6,0,0.7, ..., 0, 0.7]

Topic Modeling (LDA)
Latent Dirichlet Allocation
• You pick the number of topics
• Each topic is a distribution over words
• Each document is a distribution over topics
Easy to do in gensim
https://radimrehurek.com/gensim/models/ldamodel.html

LDA of my tweets shown in pyLDAvis
I sometimes
tweet about
movies.
https://github.com/bmabey/pyLDAvis

Sentiment analysis: How the author feels about the text
Sentiment dictionary: word list with
sentiment values associated with all words
that might indicate sentiment
...
happy: +2
...
joyful: +2
...
pain: -3
painful: -3
...
We can do sentiment analysis using labeled data and meaningless
tokens, with supervised learning over tf-idf.
We can also do sentiment analysis by
adding the first hint of meaning: some
words are positive and some words are
negative.
One such word list is VADER https://github.com/cjhutto/vaderSentiment
“I went to the junkyard and was happy to see joyful people.”

Manually Constructing Representations
We tell the computer what
things mean by manually
specifying relationships between
symbols
1. Stores meaning using
predefined relationships
2. Maps multiple ways of
writing something to the
same representation
Allows us to code
what the machine
should do for a
relatively small
number of
representations

vehicle
Honda
Civic
tires
is_ahas
Who might be in the market
for tires?
Allows us to code
what the machine
should do for a
relatively small
number of
representations
symbols
same representation

vehicle
“... the car ...”
“... an automobile ...”
“... my truck ...”
Honda
Civic
“... cruising in my civic ...”
tires
is_ahas
Who might be in the market
for tires?
Allows us to code
what the machine
should do for a
relatively small
number of
representations
symbols
same representation

WordNet
auto,
automobile,
motorcar
WordNet Search http://wordnetweb.princeton.edu/perl/webwn
ambulance
hyponym
(subordinate)
sports_car,
sport_car
hyponym
(subordinate)
motor_vehicle,
automotive_vechicle
hypernym
(superordinate)
automobile_horn,
car_horn,
motor_horn, horn,
hooter
meronym
(has-part)
• Organizes sets of words into synsets, which are meanings
• A word can be a member of more than one synset, such as “bank”
[synsets shown as boxes]

WordNet
auto,
automobile,
machine,
motorcar
WordNet Search http://wordnetweb.princeton.edu/perl/webwn
ambulance
hyponym
(subordinate)
sports_car,
sport_car
hyponym
(subordinate)
motor_vehicle,
automotive_vechicle
hypernym
(superordinate)
automobile_horn,
car_horn,
motor_horn, horn,
hooter
meronym
(has-part)
Other relations:
• has-instance
• car : Honda Civic
• has-member
• team : stringer (like “first stringer”)
• antonym
• ... a few more
• Organizes sets of words into synsets, which are meanings
• A word can be a member of more than one synset, such as “bank”

FrameNet
• More integrative than WordNet: represents situations
• One example is a child’s birthday party, another is Commerce_buy
• situations have slots (roles) that are filled
• Frames are triggered by keywords in text (more or less)
FrameNet: https://framenet.icsi.berkeley.edu/fndrupal/IntroPage
Commerce_buy: https://framenet2.icsi.berkeley.edu/fnReports/data/frameIndex.xml?frame=Commerce_buy
Commerce_buy triggered by the words: buy,
buyer, client, purchase, or purchaser
Roles: Buyer, Goods,
Money, Place (where
bought), ...
Commerce_buy
Getting
Inherits from
Commerce_buy indicates a change of possession, but we need
a world model to actually change a state.

ConceptNet
Bread
ToasterAtLocation
Automobile
RelatedTo
• Provides commonsense linkages between words
• My experience: too shallow and haphazard to be useful
Might be good for a conversation bot:
“You know what you normally find by a
toaster? Bread.”
ConceptNet: http://conceptnet.io/

YAGO: Yet Another Great Ontology
• Built on WordNet and DBpedia
• http://www.mpi-inf.mpg.de/departments/databases-and-
information-systems/research/yago-naga/yago/
• DBpedia has a machine readable page for each Wikipedia
page
• Used by IBM Watson to play Jeopardy!
• Big on named entities, like entertainers
• Browse
• https://gate.d5.mpi-inf.mpg.de/webyago3spotlx/Browser

SUMO: Suggested Upper Merged Ontology
There is also YAGO-SUMO that merges the low-
level organization of SUMO with the instance
information of YAGO.http://people.mpi-
inf.mpg.de/~gdemelo/yagosumo/
Deep: organizes concepts down to the lowest level
http://www.adampease.org/OP/
Example: cooking is a type of making that is a type
of intentional process that is a type of process that
is a physical thing that is an entity.

Image Schemas
Humans use image schemas comprehend spatial
arrangements and movements in space [Mandler, 2004]
Examples of image schemas include path, containment,
blockage, and attraction [Johnson, 1987]
Abstract concepts such as romantic relationships and
arguments are represented as metaphors to this kind of
experience [Lakoff and Johnson, 1980]
Image schemas are representations of human experience
that are common across cultures [Feldman, 2006]

Semantic Web (Linked Data)
• Broad, but not
organized or deep
• Way too complicated
• May eventually be
streamlined (e.g.
JSON-LD), and it could
be very cool if it gets
linked with deeper,
better organized data
• Tools to map text:
• FRED
• DBpedia Spotlight
• Pikes
Semantic Web Layer Cake Spaghetti Monster

World models
Computers need causal models of how the world works
and how we interact with it.
• People don’t say everything to get a message across,
just what is not covered by our shared conception
• Most efficient way to encode our shared conception is
a model
Models express how the world changes based on events
• Recall the Commerce_buy frame
• Afterward, one person has more money and another
person has less
• Read such inferences right off the model

Dimensions of models
• Probabilistic
• Deterministic compared with stochastic
• E.g., logic compared with probabilistic programming
• Factor state
• whole states compared with using variables
• E.g., finite automata compared with dynamic Bayesian networks
• Relational
• Propositional logic compared with first-order logic
• E.g., Bayesian networks compared with Markov logic networks
• Concurrent
• Model one thing compared with multiple things
• E.g., finite automata compared with Petri Nets
• Temporal
• Static compared with dynamic
• E.g., Bayesian networks compared with dynamic Bayesian
networks

Merge representations with models
Explain your conception of
conductivity?
Electrons are small spheres, and
electricity is small spheres going
through a tube. Conductivity is how
little blockage is in the tube.
Cyc has a model that uses
representations, but it is not clear if
logic is sufficiently supple.
Why does it only rain outside?
A roof blocks the path of things
from above.
The meaning of the word
“chicken” is everything explicitly
stated in the representation and
everything that can be inferred
from the world model.
“chicken”
The final step on this path is to
create a robust model around
rich representations.

meaningless
tokens
manual
representations
world models
merged
representations
symbolic
path
sub-symbolic
path
Neural networks (deep learning)
Great place to start is the
landmark publication:
Parallel Distributed Processing,
Vols. 1 and 2, Rumelhart and
McClelland, 1987

word2vec
The word2vec model learns a vector for each
word in the vocabulary.
The number of dimensions for each word vector is
the same and is usually around 300.
Unlike the tf-idf vectors, word vectors are dense,
meaning that most values are not 0.

word2vec
1. Initialize each word with a random vector
2. For each word w1 in the set of documents:
3. For each word w2 around w1:
4. Move vectors for w1 and w2 closer
together and move all others and w1
farther apart
5. Goto 2 if not done
• Skip-gram model [Mikolov et al., 2013]
• Note: there are really two vectors per word, because you don’t want a word to
be likely to be around itself, see Goldberg and
Levyhttps://arxiv.org/pdf/1402.3722v1.pdf
• First saw that double-for-loop explanation from Christopher Moody

word2vec meaning
“You shall know a word by the company it keeps.” J. R. Firth [1957]
The quote we often see:
This seems at least kind of true.
• Vectors have internal structure [Mikolov et al., 2013]
• Italy – Rome = France – Paris
• King – Queen = Man – Woman
But ... words aren’t grounded in experience; they are only
grounded in being around other words.
Can also do word2vec on ConceptNet, see https://arxiv.org/pdf/1612.03975v1.pdf

seq2seq model
The seq2seq (sequence-to-sequence) model can encode
sequences of tokens, such as sentences, into single
vectors.
It can then decode these vectors into other sequences
of tokens.
Both the encoding and decoding are done using
recurrent neural networks (RNNs).
One obvious application for this is machine translation.
For example, where the source sentences are English
and the target sentences are Spanish.

Encoding sentence meaning into a vector
h0
The
“The patient fell.”

h0
The
h1
patient

h0
The
h1
patient
h2
fell

Like a hidden Markov model, but doesn’t make the Markov
assumption and benefits from a vector representation.
h0
The
h1
patient
h2
fell
h3
.

Decoding sentence meaning
El
h3
Machine translation, or structure learning more generally.

El
h3 h4

El
h3
paciente
h4

El
h3
paciente
h4
cayó
h5
.
h5
[Cho et al., 2014]
It keeps generating until it generates a stop symbol. Note that the lengths don’t
need to be the same. It could generate the correct “se cayó.”
• Treats this task like it is devoid of meaning.
• Great that this can work on just about any kind of seq2seq
problem, but this generality highlights its limitation for use
as language understanding. No Chomsky universal grammar.

Generating image captions
Convolutional
neural network
Sad
h0
wet
h1
kid
h2
.
h3
[Karpathy and Fei-Fei, 2015]
[Vinyals et al., 2015]

Attention [Bahdanau et al., 2014]
El
h3
paciente
h4
cayó
h5
.
h5
h0
The
h1
patient
h2
fell
h3
.

Deep learning and question answering
RNNs answer questions.
What is the translation of this
phrase to French?
What is the next word?
Attention is useful for question
answering.
This can be generalized to which facts
the learner should pay attention to
when answering questions.

Bob went home.
Tim went to the junkyard.
Bob picked up the jar.
Bob went to town.
Where is the jar? A: town
• Memory Networks [Weston et al.,
2014]
• Updates memory vectors based on
a question and finds the best one to
give the output.
The office is north of the yard.
The bath is north of the office.
The yard is west of the kitchen.
How do you go from the office to
the kitchen? A: south, east
• Neural Reasoner [Peng et al.,
2015]
• Encodes the question and facts in
many layers, and the final layer is
put through a function that gives
the answer.

The network is learning linkages between sequences
of symbols, but these kinds of stories do not have
sufficiently rich linkages to our world.

External world training
If we want to talk to machines
1. We need to train them in an environment as much like our
own as possible
2. Can’t just be dialog!
Harnad [1990] http://users.ecs.soton.ac.uk/harnad/Papers/Harnad/harnad90.sgproblem.html
To understand “chicken” we need the machine
to have had as much experience with chickens
as possible.
When we say “chicken” we don’t just mean the
bird, we mean everything one can do with it
and everything it represents in our culture.“chicken”

There has been work in this direction
Industry
• OpenAI
• Universe: train on screens with VNC
• Now with Grand Theft Auto!
https://openai.com/blog/GTA-V-plus-Universe/
• Google
• Mikolov et al., A Roadmap towards
Machine Intelligence. They define an
artificial environment.
https://arxiv.org/pdf/1511.08130v2.pdf
• Facebook
• Weston, memory networks to dialogs
• Kiela et al., Virtual Embodiment: A
Scalable Long-Term Strategy for
Artificial Intelligence Res. Advocate
using video games “with a purpose.”
Academia
• Ray Mooney
• Maps text to situations
http://videolectures.net/aaai2013_moo
ney_language_learning/
• Luc Steels
• Robots come up with
vocabulary and simple grammar
• Narasimhan et al.
• Train a neural net to play text-
based adventure games
iCub

But we need more training centered in our world
Maybe if Amazon Alexa had a camera and
rotating head?
How far could we get without the benefit of a
teacher?
• Could we use eye gaze as a cue? [Yu and
Ballard, 2010]

symbolic
path
sub-symbolic
path
Building world
models based
on large, deep,
and organized
representations
Training large neural
networks in an
environment with similar
objects, relationships, and
dynamics as our own
Two paths from NLP to AI

Final thought
Open problem: What is the simplest commercially viable
task that requires commonsense knowledge and
reasoning?

Thanks for listening
Jonathan Mugan
@jmugan
www.deepgrammar.com

“chicken”

meaningless
tokens
manual
representations
merged
representations
world models
word2vec
seq2seq
question
answering
external world
training
symbolic
path
sub-symbolic
path

Data Day Seattle, From NLP to AI

Recommended

Recommended

More Related Content

What's hot

What's hot (7)

Similar to Data Day Seattle, From NLP to AI

Similar to Data Day Seattle, From NLP to AI (20)

More from Jonathan Mugan

More from Jonathan Mugan (6)

Recently uploaded

Recently uploaded (20)

Data Day Seattle, From NLP to AI

Editor's Notes